I was reading Executor just now and found that some latest changes introduced some weird code path with too much monadic chaining and unnecessary fields. I cleaned it up a bit, and also tightened up the visibility of various fields/methods. Also added some inline documentation to help understand this code better.
Author: Reynold Xin <rxin@databricks.com>
Closes#4850 from rxin/executor and squashes the following commits:
866fc60 [Reynold Xin] Code review feedback.
020efbb [Reynold Xin] Tighten up field/method visibility in Executor and made some code more clear to read.
We define and update `visitedStages` in `DAGScheduler.stageDependsOn`, but never read it. So we can safely remove it.
Author: Wenchen Fan <cloud0fan@outlook.com>
Closes#5086 from cloud-fan/minor and squashes the following commits:
24663ea [Wenchen Fan] remove un-used variable
Specifically, when calling JavaPairRDD.combineByKey(), there is a new
six-parameter method that exposes the map-side-combine boolean as the
fifth parameter and the serializer as the sixth parameter.
Author: mcheah <mcheah@palantir.com>
Closes#4634 from mccheah/pair-rdd-map-side-combine and squashes the following commits:
5c58319 [mcheah] Fixing compiler errors.
3ce7deb [mcheah] Addressing style and documentation comments.
7455c7a [mcheah] Allowing Java combineByKey to specify Serializer as well.
6ddd729 [mcheah] [SPARK-5843] Allowing map-side combine to be specified in Java.
https://issues.apache.org/jira/browse/SPARK-4012
This patch is a resubmission for https://github.com/apache/spark/pull/2864
What I am proposing in this patch is that ***when the exception is thrown from an infinite loop, we should stop the SparkContext, instead of let JVM throws exception forever***
So, in the infinite loops where we originally wrapped with a ` logUncaughtExceptions`, I changed to `tryOrStopSparkContext`, so that the Spark component is stopped
Early stopped JVM process is helpful for HA scheme design, for example,
The user has a script checking the existence of the pid of the Spark Streaming driver for monitoring the availability; with the code before this patch, the JVM process is still available but not functional when the exceptions are thrown
andrewor14, srowen , mind taking further consideration about the change?
Author: CodingCat <zhunansjtu@gmail.com>
Closes#5004 from CodingCat/SPARK-4012-1 and squashes the following commits:
589276a [CodingCat] throw fatal error again
3c72cd8 [CodingCat] address the comments
6087864 [CodingCat] revise comments
6ad3eb0 [CodingCat] stop SparkContext instead of quit the JVM process
6322959 [CodingCat] exit JVM process when the exception is thrown from an infinite loop
The current implementation include searching a HashMap many times, we can avoid this.
Actually if you look into `BlockManager.blockIdsToBlockManagers`, the core function call is [this](https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/storage/BlockManager.scala#L1258), so we can call `blockManagerMaster.getLocations` directly and avoid building a HashMap.
Author: Wenchen Fan <cloud0fan@outlook.com>
Closes#5043 from cloud-fan/small and squashes the following commits:
e959d12 [Wenchen Fan] fix style
203c493 [Wenchen Fan] some cleanup in BlockManager companion object
d409099 [Wenchen Fan] address rxin's comment
faec999 [Wenchen Fan] add regression test
2fb57aa [Wenchen Fan] imporve the getCacheLocs method
The dynamic execution code has two ways to reduce the number of executors: one
where it reduces the total number of executors it wants, by asking for an absolute
number of executors that is lower than the previous one. The second is by
explicitly killing idle executors.
YarnAllocator was mixing those up and lowering the target number of executors
when a kill was issued. Instead, trust the frontend knows what it's doing, and kill
executors without messing with other accounting. That means that if the frontend
kills an executor without lowering the target, it will get a new executor shortly.
The one situation where both actions (lower the target and kill executor) need to
happen together is when user code explicitly calls `SparkContext.killExecutors`.
In that case, issue two calls to the backend to achieve the goal.
I also did some minor cleanup in related code:
- avoid sending a request for executors when target is unchanged, to avoid log
spam in the AM
- avoid printing misleading log messages in the AM when there are no requests
to cancel
- fix a slow memory leak plus misleading error message on the driver caused by
failing to completely unregister the executor.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#5018 from vanzin/SPARK-6325 and squashes the following commits:
2e782a3 [Marcelo Vanzin] Avoid redundant logging on the AM side.
a3567cd [Marcelo Vanzin] Add parentheses.
a363926 [Marcelo Vanzin] Update logic.
a158101 [Marcelo Vanzin] [SPARK-6325] [core,yarn] Disallow reducing executor count past running count.
Author: Iulian Dragos <jaguarul@gmail.com>
Closes#5000 from dragos/issue/task-error-case and squashes the following commits:
e063627 [Iulian Dragos] Handle TASK_ERROR in Mesos scheduler backends.
ac17cf0 [Iulian Dragos] Handle missing Mesos case TASK_ERROR.
...support NFS mounts.
This is a work around for now with the goal to find a more permanent solution.
https://issues.apache.org/jira/browse/SPARK-6313
Author: nemccarthy <nathan@nemccarthy.me>
Closes#5036 from nemccarthy/master and squashes the following commits:
2eaaf42 [nemccarthy] [SPARK-6313] Update config wording doc for spark.files.useFetchCache
5de7eb4 [nemccarthy] [SPARK-6313] Add config option to disable file locks/fetchFile cache to support NFS mounts
This PR addresses a Scala compiler bug ([SI-8905](https://issues.scala-lang.org/browse/SI-8905)) that was breaking some of the Spark Java APIs. In a nutshell, it seems that methods whose implementations are inherited from generic traits sometimes have their type parameters erased to Object. This was causing methods like `DoubleRDD.min()` to throw confusing NoSuchMethodErrors at runtime.
The fix implemented here is to introduce an intermediate layer of abstract classes and inherit from those instead of directly extends the `Java*Like` traits. This should not break binary compatibility.
I also improved the test coverage of the Java API, adding several new tests for methods that failed at runtime due to this bug.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#5050 from JoshRosen/javardd-si-8905-fix and squashes the following commits:
2feb068 [Josh Rosen] Use intermediate abstract classes to work around SPARK-3266
d5f3e5d [Josh Rosen] Add failing regression tests for SPARK-3266
This patch switches the usage of java.lang.reflect.Array in Size estimation to using scala's RunTime array-getter methods. The notes on https://bugs.openjdk.java.net/browse/JDK-8051447 tipped me off to the fact that using java.lang.reflect.Array was not ideal. At first, I used the code from that ticket, but it turns out that ScalaRunTime's array-related methods avoid the bottleneck of invoking native code anyways, so that was sufficient to boost performance in size estimation.
The idea is to use pure Java code in implementing the methods there, as opposed to relying on native C code which ends up being ill-performing. This improves the performance of estimating the size of arrays when we are checking for spilling in Spark.
Here's the benchmark discussion from the ticket:
I did two tests. The first, less convincing, take-with-a-block-of-salt test I did was do a simple groupByKey operation to collect objects in a 4.0 GB text file RDD into 30,000 buckets. I ran 1 Master and 4 Spark Worker JVMs on my mac, fetching the RDD from a text file simply stored on disk, and saving it out to another file located on local disk. The wall clock times I got back before and after the change were:
Before: 352.195s, 343.871s, 359.080s
After (using code directly from the JDK ticket, not the scala code in this PR): 342.929583s, 329.456623s, 326.151481s
So, there is a bit of an improvement after the change. I also did some YourKit profiling of the executors to get an idea of how much time was spent in size estimation before and after the change. I roughly saw that size estimation took up less of the time after my change, but YourKit's profiling can be inconsistent and who knows if I was profiling the executors that had the same data between runs?
The more convincing test I did was to run the size-estimation logic itself in an isolated unit test. I ran the following code:
```
val bigArray = Array.fill(1000)(Array.fill(1000)(java.util.UUID.randomUUID().toString()))
test("String arrays only perf testing") {
val startTime = System.currentTimeMillis()
for (i <- 1 to 50000) {
SizeEstimator.estimate(bigArray)
}
println("Runtime: " + (System.currentTimeMillis() - startTime) / 1000.0000)
}
```
I wanted to use a 2D array specifically because I wanted to measure the performance of repeatedly calling Array.getLength. I used UUID-Strings to ensure that the strings were randomized (so String object re-use doesn't happen), but that they would all be the same size. The results were as follows:
Before PR: 222.681 s, 218.34 s, 211.739s
After latest change: 170.715 s, 176.775 s, 180.298 s
.
Author: mcheah <mcheah@palantir.com>
Author: Justin Uang <justin.uang@gmail.com>
Closes#4972 from mccheah/feature/spark-6269-reflect-array and squashes the following commits:
8527852 [mcheah] Respect CamelCase for numElementsDrawn
18d4b50 [mcheah] Addressing style comments - while loops instead of for loops
16ce534 [mcheah] Organizing imports properly
db890ea [mcheah] Removing CastedArray and just using ScalaRunTime.
cb67ce2 [mcheah] Fixing a scalastyle error - line too long
5d53c4c [mcheah] Removing unused parameter in visitArray.
6467759 [mcheah] Including primitive size information inside CastedArray.
93f4b05 [mcheah] Using Scala instead of Java for the array-reflection implementation.
a557ab8 [mcheah] Using a wrapper around arrays to do casting only once
ca063fc [mcheah] Fixing a compiler error made while refactoring style
1fe09de [Justin Uang] [SPARK-6269] Use a different implementation of java.lang.reflect.Array
https://issues.apache.org/jira/browse/SPARK-4011
Currently, most of the members in Master/Worker are with public accessibility. We might wish to tighten the accessibility of them
a bit more discussion is here:
https://github.com/apache/spark/pull/2828
Author: CodingCat <zhunansjtu@gmail.com>
Closes#4844 from CodingCat/SPARK-4011 and squashes the following commits:
1a64175 [CodingCat] fix compilation issue
e7fd375 [CodingCat] Sean is right....
f5034a4 [CodingCat] fix rebase mistake
8d5b0c0 [CodingCat] loose more fields
0072f96 [CodingCat] lose some restrictions based on the possible design intention
de77286 [CodingCat] tighten accessibility of deploy package
12b4fd3 [CodingCat] tighten accessibility of deploy.worker
1243bc7 [CodingCat] tighten accessibility of deploy.rest
c5f622c [CodingCat] tighten the accessibility of deploy.history
d441e20 [CodingCat] tighten accessibility of deploy.client
4e0ce4a [CodingCat] tighten the accessibility of the members of classes in master
23cddbb [CodingCat] stylistic fix
9a3a340 [CodingCat] tighten the access of worker class
67a0559 [CodingCat] tighten the access permission in Master
Use configured serializer in RDD.aggregate, to match PairRDDFunctions.aggregateByKey, instead of closure serializer.
Compare with e60ad2f4c4/core/src/main/scala/org/apache/spark/rdd/PairRDDFunctions.scala (L127)
Author: Sean Owen <sowen@cloudera.com>
Closes#5028 from srowen/SPARK-6044 and squashes the following commits:
a4040a7 [Sean Owen] Use configured serializer in RDD.aggregate, to match PairRDDFunctions.aggregateByKey, instead of closure serializer
```
case class ClassA(value: String)
val rdd = sc.parallelize(List(("k1", ClassA("v1")), ("k1", ClassA("v2")) ))
rdd.groupByKey.collect
```
This code used to be throw exception in spark-shell, because while shuffling ```JavaSerializer```uses ```defaultClassLoader``` which was defined like ```env.serializer.setDefaultClassLoader(urlClassLoader)```.
It should be ```env.serializer.setDefaultClassLoader(replClassLoader)```, like
```
override def run() {
val deserializeStartTime = System.currentTimeMillis()
Thread.currentThread.setContextClassLoader(replClassLoader)
```
in TaskRunner.
When ```replClassLoader``` cannot be defined, it's identical with ```urlClassLoader```
Author: Kevin (Sangwoo) Kim <sangwookim.me@gmail.com>
Closes#5046 from swkimme/master and squashes the following commits:
fa2b9ee [Kevin (Sangwoo) Kim] stylish test codes ( collect -> collect() )
6e9620b [Kevin (Sangwoo) Kim] stylish test codes ( collect -> collect() )
d23e4e2 [Kevin (Sangwoo) Kim] stylish test codes ( collect -> collect() )
a4a3c8a [Kevin (Sangwoo) Kim] add 'class defined in repl - shuffle' test to ReplSuite
bd00da5 [Kevin (Sangwoo) Kim] add 'class defined in repl - shuffle' test to ReplSuite
c1b1fc7 [Kevin (Sangwoo) Kim] use REPL class loader for executor's serializer
Currently we would create a new streaming tab for each streamingContext even if there's already one on the same sparkContext which would cause duplicate StreamingTab created and none of them is taking effect.
snapshot: https://www.dropbox.com/s/t4gd6hqyqo0nivz/bad%20multiple%20streamings.png?dl=0
How to reproduce:
1)
import org.apache.spark.SparkConf
import org.apache.spark.streaming.
{Seconds, StreamingContext}
import org.apache.spark.storage.StorageLevel
val ssc = new StreamingContext(sc, Seconds(1))
val lines = ssc.socketTextStream("localhost", 9999, StorageLevel.MEMORY_AND_DISK_SER)
val words = lines.flatMap(_.split(" "))
val wordCounts = words.map(x => (x, 1)).reduceByKey(_ + _)
wordCounts.print()
ssc.start()
.....
2)
ssc.stop(false)
val ssc = new StreamingContext(sc, Seconds(1))
val lines = ssc.socketTextStream("localhost", 9999, StorageLevel.MEMORY_AND_DISK_SER)
val words = lines.flatMap(_.split(" "))
val wordCounts = words.map(x => (x, 1)).reduceByKey(_ + _)
wordCounts.print()
ssc.start()
Author: lisurprise <zhichao.li@intel.com>
Closes#4828 from zhichao-li/master and squashes the following commits:
c329806 [lisurprise] add test for attaching/detaching streaming tab
51e6c7f [lisurprise] move detach method into StreamingTab
31a44fa [lisurprise] add unit test for attaching and detaching new tab
db25ed2 [lisurprise] clean code
8281bcb [lisurprise] clean code
193c542 [lisurprise] remove streaming tab while closing streaming context
when i run cmd like that sc.addFile("../test.txt"), it did not work and throwed an exception:
java.lang.IllegalArgumentException: java.net.URISyntaxException: Relative path in absolute URI: file:../test.txt
at org.apache.hadoop.fs.Path.initialize(Path.java:206)
at org.apache.hadoop.fs.Path.<init>(Path.java:172)
........
.......
Caused by: java.net.URISyntaxException: Relative path in absolute URI: file:../test.txt
at java.net.URI.checkPath(URI.java:1804)
at java.net.URI.<init>(URI.java:752)
at org.apache.hadoop.fs.Path.initialize(Path.java:203)
Author: DoingDone9 <799203320@qq.com>
Closes#4993 from DoingDone9/relativePath and squashes the following commits:
ee375cd [DoingDone9] Update SparkContextSuite.scala
d594e16 [DoingDone9] Update SparkContext.scala
0ff3fa8 [DoingDone9] test for add file
dced8eb [DoingDone9] Update SparkContext.scala
e4a13fe [DoingDone9] getCanonicalPath
161cae3 [DoingDone9] Merge pull request #4 from apache/master
c87e8b6 [DoingDone9] Merge pull request #3 from apache/master
cb1852d [DoingDone9] Merge pull request #2 from apache/master
c3f046f [DoingDone9] Merge pull request #1 from apache/master
This existed at the very beginning, but became unnecessary after [this commit](37d8f37a8e (diff-6a9ff7fb74fd490a50462d45db2d5e11L272)). I think we should remove it if we don't plan to use it in the future.
Author: Wenchen Fan <cloud0fan@outlook.com>
Closes#4992 from cloud-fan/small and squashes the following commits:
e857f2e [Wenchen Fan] remove unnecessary ClassTag
For details, please refer to [SPARK-6197](https://issues.apache.org/jira/browse/SPARK-6197)
Author: Zhang, Liye <liye.zhang@intel.com>
Closes#4927 from liyezhang556520/jsonParseError and squashes the following commits:
5cbdc82 [Zhang, Liye] without unnecessary wrap
2b48831 [Zhang, Liye] small changes with sean owen's comments
2973024 [Zhang, Liye] handle json exception when file not finished writing
I've added a timer in the right place to fix this inaccuracy.
Author: Ilya Ganelin <ilya.ganelin@capitalone.com>
Closes#4965 from ilganeli/SPARK-5845 and squashes the following commits:
bfabf88 [Ilya Ganelin] Changed to using a foreach vs. getorelse
3e059b0 [Ilya Ganelin] Switched to using getorelse
b946d08 [Ilya Ganelin] Fixed error with option
9434b50 [Ilya Ganelin] Merge remote-tracking branch 'upstream/master' into SPARK-5845
db8647e [Ilya Ganelin] Added update for shuffleWriteTime around spilled file cleanup in ExternalSorter
The Thread.interrupt() can not terminate the thread in some cases, so we should not wait for the writerThread of PythonRDD.
This PR also ignore some exception during clean up.
cc JoshRosen mengxr
Author: Davies Liu <davies@databricks.com>
Closes#4987 from davies/fix_take and squashes the following commits:
4488f1a [Davies Liu] fix hang when call take() in JVM on PythonRDD
Resolve javac, scalac warnings of various types -- deprecations, Scala lang, unchecked cast, etc.
Author: Sean Owen <sowen@cloudera.com>
Closes#4950 from srowen/SPARK-6225 and squashes the following commits:
3080972 [Sean Owen] Ordered imports: Java, Scala, 3rd party, Spark
c67985b [Sean Owen] Resolve javac, scalac warnings of various types -- deprecations, Scala lang, unchecked cast, etc.
This change encapsulates all the logic involved in launching a Spark job
into a small Java library that can be easily embedded into other applications.
The overall goal of this change is twofold, as described in the bug:
- Provide a public API for launching Spark processes. This is a common request
from users and currently there's no good answer for it.
- Remove a lot of the duplicated code and other coupling that exists in the
different parts of Spark that deal with launching processes.
A lot of the duplication was due to different code needed to build an
application's classpath (and the bootstrapper needed to run the driver in
certain situations), and also different code needed to parse spark-submit
command line options in different contexts. The change centralizes those
as much as possible so that all code paths can rely on the library for
handling those appropriately.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#3916 from vanzin/SPARK-4924 and squashes the following commits:
18c7e4d [Marcelo Vanzin] Fix make-distribution.sh.
2ce741f [Marcelo Vanzin] Add lots of quotes.
3b28a75 [Marcelo Vanzin] Update new pom.
a1b8af1 [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
897141f [Marcelo Vanzin] Review feedback.
e2367d2 [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
28cd35e [Marcelo Vanzin] Remove stale comment.
b1d86b0 [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
00505f9 [Marcelo Vanzin] Add blurb about new API in the programming guide.
5f4ddcc [Marcelo Vanzin] Better usage messages.
92a9cfb [Marcelo Vanzin] Fix Win32 launcher, usage.
6184c07 [Marcelo Vanzin] Rename field.
4c19196 [Marcelo Vanzin] Update comment.
7e66c18 [Marcelo Vanzin] Fix pyspark tests.
0031a8e [Marcelo Vanzin] Review feedback.
c12d84b [Marcelo Vanzin] Review feedback. And fix spark-submit on Windows.
e2d4d71 [Marcelo Vanzin] Simplify some code used to launch pyspark.
43008a7 [Marcelo Vanzin] Don't make builder extend SparkLauncher.
b4d6912 [Marcelo Vanzin] Use spark-submit script in SparkLauncher.
28b1434 [Marcelo Vanzin] Add a comment.
304333a [Marcelo Vanzin] Fix propagation of properties file arg.
bb67b93 [Marcelo Vanzin] Remove unrelated Yarn change (that is also wrong).
8ec0243 [Marcelo Vanzin] Add missing newline.
95ddfa8 [Marcelo Vanzin] Fix handling of --help for spark-class command builder.
72da7ec [Marcelo Vanzin] Rename SparkClassLauncher.
62978e4 [Marcelo Vanzin] Minor cleanup of Windows code path.
9cd5b44 [Marcelo Vanzin] Make all non-public APIs package-private.
e4c80b6 [Marcelo Vanzin] Reorganize the code so that only SparkLauncher is public.
e50dc5e [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
de81da2 [Marcelo Vanzin] Fix CommandUtils.
86a87bf [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
2061967 [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
46d46da [Marcelo Vanzin] Clean up a test and make it more future-proof.
b93692a [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
ad03c48 [Marcelo Vanzin] Revert "Fix a thread-safety issue in "local" mode."
0b509d0 [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
23aa2a9 [Marcelo Vanzin] Read java-opts from conf dir, not spark home.
7cff919 [Marcelo Vanzin] Javadoc updates.
eae4d8e [Marcelo Vanzin] Fix new unit tests on Windows.
e570fb5 [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
44cd5f7 [Marcelo Vanzin] Add package-info.java, clean up javadocs.
f7cacff [Marcelo Vanzin] Remove "launch Spark in new thread" feature.
7ed8859 [Marcelo Vanzin] Some more feedback.
54cd4fd [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
61919df [Marcelo Vanzin] Clean leftover debug statement.
aae5897 [Marcelo Vanzin] Use launcher classes instead of jars in non-release mode.
e584fc3 [Marcelo Vanzin] Rework command building a little bit.
525ef5b [Marcelo Vanzin] Rework Unix spark-class to handle argument with newlines.
8ac4e92 [Marcelo Vanzin] Minor test cleanup.
e946a99 [Marcelo Vanzin] Merge PySparkLauncher into SparkSubmitCliLauncher.
c617539 [Marcelo Vanzin] Review feedback round 1.
fc6a3e2 [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
f26556b [Marcelo Vanzin] Fix a thread-safety issue in "local" mode.
2f4e8b4 [Marcelo Vanzin] Changes needed to make this work with SPARK-4048.
799fc20 [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
bb5d324 [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
53faef1 [Marcelo Vanzin] Merge branch 'master' into SPARK-4924
a7936ef [Marcelo Vanzin] Fix pyspark tests.
656374e [Marcelo Vanzin] Mima fixes.
4d511e7 [Marcelo Vanzin] Fix tools search code.
7a01e4a [Marcelo Vanzin] Fix pyspark on Yarn.
1b3f6e9 [Marcelo Vanzin] Call SparkSubmit from spark-class launcher for unknown classes.
25c5ae6 [Marcelo Vanzin] Centralize SparkSubmit command line parsing.
27be98a [Marcelo Vanzin] Modify Spark to use launcher lib.
6f70eea [Marcelo Vanzin] [SPARK-4924] Add a library for launching Spark jobs programatically.
A simple try-catch wrapping KryoException to be more informative.
Author: Lev Khomich <levkhomich@gmail.com>
Closes#4947 from levkhomich/master and squashes the following commits:
0f7a947 [Lev Khomich] [SPARK-6087][CORE] Provide actionable exception if Kryo buffer is not large enough
Because circular reference between JavaObject and JavaMember, an Java object can not be released until Python GC kick in, then it will cause memory leak in collect(), which may consume lots of memory in JVM.
This PR change the way we sending collected data back into Python from local file to socket, which could avoid any disk IO during collect, also avoid any referrers of Java object in Python.
cc JoshRosen
Author: Davies Liu <davies@databricks.com>
Closes#4923 from davies/fix_collect and squashes the following commits:
d730286 [Davies Liu] address comments
24c92a4 [Davies Liu] fix style
ba54614 [Davies Liu] use socket to transfer data from JVM
9517c8f [Davies Liu] fix memory leak in collect()
Resolve sbt warnings:
```
[warn] spark/streaming/src/main/scala/org/apache/spark/streaming/util/WriteAheadLogManager.scala:155: postfix operator second should be enabled
[warn] by making the implicit value scala.language.postfixOps visible.
[warn] This can be achieved by adding the import clause 'import scala.language.postfixOps'
[warn] or by setting the compiler option -language:postfixOps.
[warn] See the Scala docs for value scala.language.postfixOps for a discussion
[warn] why the feature should be explicitly enabled.
[warn] Await.ready(f, 1 second)
[warn] ^
```
Author: GuoQiang Li <witgo@qq.com>
Closes#4908 from witgo/sbt_warnings and squashes the following commits:
0629af4 [GuoQiang Li] Resolve sbt warnings: postfix operator second should be enabled
This patch fixes two issues with the executor log viewing links added in Spark 1.3. In standalone mode, the log URLs might include a port value of 0 rather than the actual bound port of the UI, which broke the ability to view logs from workers whose web UIs had been configured to bind to ephemeral ports. In addition, the URLs used workers' local hostnames instead of respecting SPARK_PUBLIC_DNS, which prevented this feature from working properly on Spark EC2 clusters because the links would point to internal DNS names instead of external ones.
I included tests for both of these bugs:
- We now browse to the URLs and verify that they point to the expected pages.
- To test SPARK_PUBLIC_DNS, I changed the code that reads the environment variable to do so via `SparkConf.getenv`, then used a custom SparkConf subclass to mock the environment variable (this pattern is used elsewhere in Spark's tests).
Author: Josh Rosen <joshrosen@databricks.com>
Closes#4903 from JoshRosen/SPARK-6175 and squashes the following commits:
5577f41 [Josh Rosen] Remove println
cfec135 [Josh Rosen] Use webUi.boundPort and publicAddress in log links
27918c7 [Josh Rosen] Add failing unit tests for standalone log URL viewing
c250fbe [Josh Rosen] Respect SparkConf in local-cluster Workers.
422a2ef [Josh Rosen] Use conf.getenv to read SPARK_PUBLIC_DNS
The code failed in two modes: it complained when it tried to re-create a directory that already existed, and it was placing some files in the wrong parent directory. The patch fixes both issues.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Author: trystanleftwich <trystan@atscale.com>
Closes#4894 from vanzin/SPARK-6144 and squashes the following commits:
100b3a1 [Marcelo Vanzin] Style fix.
58266aa [Marcelo Vanzin] Fix fetchHcfs file for directories.
91733b7 [trystanleftwich] [SPARK-6144]When in cluster mode using ADD JAR with a hdfs:// sourced jar will fail
when application is finished running abnormally (Ctrl + c for example), the history event log file is still ends with `.inprogress` suffix. And the application state can not be showed on webUI, User can only see "*Application history not foud xxxx, Application xxx is still in progress*".
For application that not finished normally, the history will show:
Author: Zhang, Liye <liye.zhang@intel.com>
Closes#4848 from liyezhang556520/showLogInprogress and squashes the following commits:
03589ac [Zhang, Liye] change inprogress to in progress
b55f19f [Zhang, Liye] scala modify after rebase
8aa66a2 [Zhang, Liye] use softer wording
b030bd4 [Zhang, Liye] clean code
79c8cb1 [Zhang, Liye] fix some mistakes
11cdb68 [Zhang, Liye] add a missing space
c29205b [Zhang, Liye] refine code according to sean owen's comments
e9952a7 [Zhang, Liye] scala style fix again
150502d [Zhang, Liye] scala style fix
f11a5da [Zhang, Liye] small fix for file path
22e878b [Zhang, Liye] enable in progress eventlog file
Author: tedyu <yuzhihong@gmail.com>
Closes#4836 from tedyu/master and squashes the following commits:
d65b495 [tedyu] SPARK-6085 Increase default value for memory overhead
1fdd4df [tedyu] SPARK-6085 Increase default value for memory overhead
If the cleaner is stopped, we shouldn't print a huge stack trace when the cleaner thread is interrupted because we purposefully did this.
Author: Andrew Or <andrew@databricks.com>
Closes#4882 from andrewor14/cleaner-interrupt and squashes the following commits:
8652120 [Andrew Or] Just a hot fix
Before we would get the following (benign) error if we called `sc.stop()` twice. This is because the listener bus would try to post the end event again even after it has already stopped. This happens occasionally when flaky tests fail, usually as a result of other sources of error. Either way we shouldn't be logging this error when it is not the cause of the failure.
```
ERROR LiveListenerBus: SparkListenerBus has already stopped! Dropping event SparkListenerApplicationEnd(1425348445682)
```
Author: Andrew Or <andrew@databricks.com>
Closes#4871 from andrewor14/sc-stop and squashes the following commits:
a14afc5 [Andrew Or] Move code after code
915db16 [Andrew Or] Move code into code
The problem is that `ContextCleaner` may clean variables that belong to a different `SparkContext`. This can happen if the `SparkContext` to which the cleaner belongs stops, and a new one is started immediately afterwards in the same JVM. In this case, if the cleaner is in the middle of cleaning a broadcast, for instance, it will do so through `SparkEnv.get.blockManager`, which could be one that belongs to a different `SparkContext`.
JoshRosen and I suspect that this is the cause of many flaky tests, most notably the `JavaAPISuite`. We were able to reproduce the failure locally (though it is not deterministic and very hard to reproduce).
Author: Andrew Or <andrew@databricks.com>
Closes#4869 from andrewor14/cleaner-masquerade and squashes the following commits:
29168c0 [Andrew Or] Synchronize ContextCleaner stop
Use array copy instead of `Iterator#toArray` to make it more efficient.
Author: Wenchen Fan <cloud0fan@outlook.com>
Closes#4825 from cloud-fan/minor and squashes the following commits:
c933ee5 [Wenchen Fan] make toArray method just in parent
946a35b [Wenchen Fan] minor enhance
Some memory not count into memory used by memoryStore or unrollMemory.
Thread A after unrollsafely memory, it will release 40MB unrollMemory(40MB will used by other threads). then ThreadA wait get accountingLock to tryToPut blockA(30MB). before Thread A get accountingLock, blockA memory size is not counting into unrollMemory or memoryStore.currentMemory.
IIUC, freeMemory should minus that block memory
So, put this release memory into pending, and release it in tryToPut before ensureSpace
Author: hushan[胡珊] <hushan@xiaomi.com>
Closes#3629 from suyanNone/unroll-memory and squashes the following commits:
809cc41 [hushan[胡珊]] Refine
407b2c9 [hushan[胡珊]] Refine according comments
39960d0 [hushan[胡珊]] Refine comments
0fd0213 [hushan[胡珊]] add comments
0fc2bec [hushan[胡珊]] Release pending unroll memory after put block in memoryStore
3a3f2c8 [hushan[胡珊]] Refine blockManagerSuite unroll test
3323c45 [hushan[胡珊]] Refine getOrElse
f664317 [hushan[胡珊]] Make sure not add pending in every releaseUnrollMemory call
08b32ba [hushan[胡珊]] Pending unroll memory for this block untill tryToPut
There are multiple issues with translating on set outlined in the JIRA.
This PR reverts the translation logic added to `SparkConf`. In the future, after the 1.3.0 release we will figure out a way to reorganize the internal structure more elegantly. For now, let's preserve the existing semantics of `SparkConf` since it's a public interface. Unfortunately this means duplicating some code for now, but this is all internal and we can always clean it up later.
Author: Andrew Or <andrew@databricks.com>
Closes#4799 from andrewor14/conf-set-translate and squashes the following commits:
11c525b [Andrew Or] Move warning to driver
10e77b5 [Andrew Or] Add documentation for deprecation precedence
a369cb1 [Andrew Or] Merge branch 'master' of github.com:apache/spark into conf-set-translate
c26a9e3 [Andrew Or] Revert all translate logic in SparkConf
fef6c9c [Andrew Or] Restore deprecation logic for spark.executor.userClassPathFirst
94b4dfa [Andrew Or] Translate on get, not set
Some users have reported difficulty in parsing the new event log format. Since we embed the metadata in the beginning of the file, when we compress the event log we need to skip the metadata because we need that information to parse the log later. This means we'll end up with a partially compressed file if event logging compression is turned on. The old format looks like:
```
sparkVersion = 1.3.0
compressionCodec = org.apache.spark.io.LZFCompressionCodec
=== LOG_HEADER_END ===
// actual events, could be compressed bytes
```
The new format in this patch puts the compression codec in the log file name instead. It also removes the metadata header altogether along with the Spark version, which was not needed. The new file name looks something like:
```
app_without_compression
app_123.lzf
app_456.snappy
```
I tested this with and without compression, using different compression codecs and event logging directories. I verified that both the `Master` and the `HistoryServer` can render both compressed and uncompressed logs as before.
Author: Andrew Or <andrew@databricks.com>
Closes#4821 from andrewor14/event-log-format and squashes the following commits:
8511141 [Andrew Or] Fix test
654883d [Andrew Or] Add back metadata with Spark version
7f537cd [Andrew Or] Address review feedback
7d6aa61 [Andrew Or] Make codec an extension
59abee9 [Andrew Or] Merge branch 'master' of github.com:apache/spark into event-log-format
27c9a6c [Andrew Or] Address review feedback
519e51a [Andrew Or] Address review feedback
ef69276 [Andrew Or] Merge branch 'master' of github.com:apache/spark into event-log-format
88a091d [Andrew Or] Add tests for new format and file name
f32d8d2 [Andrew Or] Fix tests
8db5a06 [Andrew Or] Embed metadata in the event log file name instead
When starting the history server, all the log files will be fetched and parsed in order to get the applications' meta data e.g. App Name, Start Time, Duration, etc. In our production cluster, there exist 2600 log files (160G) in HDFS and it costs 3 hours to restart the history server, which is a little bit too long for us.
It would be better, if the history server can show logs with missing information during start-up and fill the missing information after fetching and parsing a log file.
Author: guliangliang <guliangliang@qiyi.com>
Closes#4525 from marsishandsome/Spark5522 and squashes the following commits:
a865c11 [guliangliang] fix bug2
4340c2b [guliangliang] fix bug
af92a5a [guliangliang] [SPARK-5522] Accelerate the Histroty Server start
Demote some 'noisy' log messages to debug level. I added a few more, to include everything that gets logged in stanzas like this:
```
15/03/01 00:03:54 INFO BlockManager: Removing broadcast 0
15/03/01 00:03:54 INFO BlockManager: Removing block broadcast_0_piece0
15/03/01 00:03:54 INFO MemoryStore: Block broadcast_0_piece0 of size 839 dropped from memory (free 277976091)
15/03/01 00:03:54 INFO BlockManagerInfo: Removed broadcast_0_piece0 on localhost:49524 in memory (size: 839.0 B, free: 265.1 MB)
15/03/01 00:03:54 INFO BlockManagerMaster: Updated info of block broadcast_0_piece0
15/03/01 00:03:54 INFO BlockManager: Removing block broadcast_0
15/03/01 00:03:54 INFO MemoryStore: Block broadcast_0 of size 1088 dropped from memory (free 277977179)
15/03/01 00:03:54 INFO ContextCleaner: Cleaned broadcast 0
```
as well as regular messages like
```
15/03/01 00:02:33 INFO MemoryStore: ensureFreeSpace(2640) called with curMem=47322, maxMem=278019440
```
WDYT? good or should some be left alone?
CC mengxr who suggested some of this.
Author: Sean Owen <sowen@cloudera.com>
Closes#4838 from srowen/SPARK-3357 and squashes the following commits:
dce75c1 [Sean Owen] Back out some debug level changes
d9b784d [Sean Owen] Demote some 'noisy' log messages to debug level
This fixes a non-deterministic bug introduced in #4021 that could cause tasks' accumulator updates to be lost. The problem is that `localAccums` should not hold weak references: after the task finishes running there won't be any strong references to these local accumulators, so they can get garbage-collected before the executor reads the `localAccums` map. We don't need weak references here anyways, since this map is cleared at the end of each task.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#4835 from JoshRosen/SPARK-6075 and squashes the following commits:
4f4b5b2 [Josh Rosen] Remove defensive assertions that caused test failures in code unrelated to this change
120c7b0 [Josh Rosen] [SPARK-6075] Do not store WeakReferences in localAccums map
Don't throw NPE if appId is unknown. kayousterhout is this a decent enough band-aid for avoiding a full-blown NPE? it should just render empty content instead
Author: Sean Owen <sowen@cloudera.com>
Closes#4777 from srowen/SPARK-1965 and squashes the following commits:
7e16590 [Sean Owen] Update app not found message
cb878d6 [Sean Owen] Return basic "not found" page for unknown appId
d8270da [Sean Owen] Don't throw NPE if appId is unknown
Disallow TRACE HTTP method in servlets
Author: Sean Owen <sowen@cloudera.com>
Closes#4765 from srowen/SPARK-5983 and squashes the following commits:
421b25b [Sean Owen] Disallow TRACE HTTP method in servlets
pwendell tdas
This is the safer parts of PR #4754:
- SPARK-5979: All dependencies with the groupId `org.apache.spark` passed through `--packages`, were being excluded from the dependency tree on the assumption that they would be in the assembly jar. This is not the case, therefore the exclusion rules had to be defined more explicitly.
- SPARK-6032: Ivy prints a whole lot of logs while retrieving dependencies. These were printed to `System.out`. Moved the logging to `System.err`.
Author: Burak Yavuz <brkyvz@gmail.com>
Closes#4802 from brkyvz/simple-streaming-fix and squashes the following commits:
e0f38cb [Burak Yavuz] Merge branch 'master' of github.com:apache/spark into simple-streaming-fix
bad921c [Burak Yavuz] [SPARK-5979][SPARK-6032] Smaller safer fix
For detail description, pls refer to [SPARK-6036](https://issues.apache.org/jira/browse/SPARK-6036).
Author: Zhang, Liye <liye.zhang@intel.com>
Closes#4785 from liyezhang556520/EventLogInProcess and squashes the following commits:
8b0b0a6 [Zhang, Liye] stop listener after DAGScheduler
79b15b3 [Zhang, Liye] SPARK-6036 avoid race condition between eventlogListener and akka actor system
The warning of deprecated configs is actually done when the configs are set, not when they are get. As a result we don't need to explicitly call `translateConfKey` outside of `SparkConf` just to print the warning again in vain.
Author: Andrew Or <andrew@databricks.com>
Closes#4797 from andrewor14/warn-deprecated-config and squashes the following commits:
8fb43e6 [Andrew Or] Privatize SparkConf.translateConfKey
Add application kill function in master web UI for standalone mode. Details can be seen in [SPARK-5495](https://issues.apache.org/jira/browse/SPARK-5495).
The snapshot of UI shows as below:
Please help to review, thanks a lot.
Author: jerryshao <saisai.shao@intel.com>
Closes#4288 from jerryshao/SPARK-5495 and squashes the following commits:
fa3e486 [jerryshao] Add some conditions
9a7be93 [jerryshao] Add kill Driver function
a239776 [jerryshao] Change the code format
ff5195d [jerryshao] Add app kill function in master web UI
cc andrewor14, srowen.
Author: jerryshao <saisai.shao@intel.com>
Closes#4800 from jerryshao/SPARK-5771 and squashes the following commits:
a2483c2 [jerryshao] Change the UI of Requested Cores into * if default cores is not set
If a blockManager has not send heartBeat more than 120s, BlockManagerMasterActor will remove it. But coarseGrainedSchedulerBackend can only remove executor after an DisassociatedEvent. We should expireDeadHosts at HeartbeatReceiver.
Author: Hong Shen <hongshen@tencent.com>
Closes#4363 from shenh062326/my_change3 and squashes the following commits:
2c9a46a [Hong Shen] Change some code style.
1a042ff [Hong Shen] Change some code style.
2dc456e [Hong Shen] Change some code style.
d221493 [Hong Shen] Fix test failed
7448ac6 [Hong Shen] A minor change in sparkContext and heartbeatReceiver
b904aed [Hong Shen] Fix failed test
52725af [Hong Shen] Remove assert in SparkContext.killExecutors
5bedcb8 [Hong Shen] Remove assert in SparkContext.killExecutors
a858fb5 [Hong Shen] A minor change in HeartbeatReceiver
3e221d9 [Hong Shen] A minor change in HeartbeatReceiver
6bab7aa [Hong Shen] Change a code style.
07952f3 [Hong Shen] Change configs name and code style.
ce9257e [Hong Shen] Fix test failed
bccd515 [Hong Shen] Fix test failed
8e77408 [Hong Shen] Fix test failed
c1dfda1 [Hong Shen] Fix test failed
e197e20 [Hong Shen] Fix test failed
fb5df97 [Hong Shen] Remove ExpireDeadHosts in BlockManagerMessages
b5c0441 [Hong Shen] Remove expireDeadHosts in BlockManagerMasterActor
c922cb0 [Hong Shen] Add expireDeadHosts in HeartbeatReceiver
Ensure scheduler delay handles unfinished task case, and ensure delay is never negative even due to rounding
Author: Sean Owen <sowen@cloudera.com>
Closes#4796 from srowen/SPARK-4579 and squashes the following commits:
ad6713c [Sean Owen] Ensure scheduler delay handles unfinished task case, and ensure delay is never negative even due to rounding
...ns#saveAsNewAPIHadoopDataset
Author: tedyu <yuzhihong@gmail.com>
Closes#4794 from tedyu/master and squashes the following commits:
2632a57 [tedyu] SPARK-6045 RecordWriter should be checked against null in PairRDDFunctions#saveAsNewAPIHadoopDataset
2d8d4b1 [tedyu] SPARK-6045 RecordWriter should be checked against null in PairRDDFunctions#saveAsNewAPIHadoopDataset
Close appender saving stdout/stderr before destroying process to avoid exception on reading closed input stream.
(This also removes a redundant `waitFor()` although it was harmless)
CC tdas since I think you wrote this method.
Author: Sean Owen <sowen@cloudera.com>
Closes#4787 from srowen/SPARK-4300 and squashes the following commits:
e0cdabf [Sean Owen] Close appender saving stdout/stderr before destroying process to avoid exception on reading closed input stream
Join on output threads to make sure any lingering output from process reaches stdout, stderr before exiting
CC andrewor14 since I believe he created this section of code
Author: Sean Owen <sowen@cloudera.com>
Closes#4788 from srowen/SPARK-4704 and squashes the following commits:
ad7114e [Sean Owen] Join on output threads to make sure any lingering output from process reaches stdout, stderr before exiting
Removing elements from a mutable HashSet while iterating over it can cause the
iteration to incorrectly skip over entries that were not removed. If this
happened, PythonRDD would write fewer broadcast variables than the Python
worker was expecting to read, which would cause the Python worker to hang
indefinitely.
Author: Davies Liu <davies@databricks.com>
Closes#4776 from davies/fix_hang and squashes the following commits:
a4384a5 [Davies Liu] fix bug: remvoe() inside iterator is not safe
Cache the value of the local root dirs to use for storing local data,
so that the same directories are reused.
Also, to avoid an extra level of nesting, use a different env variable
to propagate the local dirs from the Worker to the executors. And make
the executor directory use a different name.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#4747 from vanzin/SPARK-5801 and squashes the following commits:
e0114e1 [Marcelo Vanzin] Update unit test.
18ee0a7 [Marcelo Vanzin] [SPARK-5801] [core] Avoid creating nested directories.
Because windows on-default does not grant read permission to jars except to admin, spark-submit would fail with "ClassNotFound" exception if user runs slave service with only user permission.
This fix is to add read permission to owner of the jar (which would be the slave service account in windows )
Author: Judy Nash <judynash@microsoft.com>
Closes#4742 from judynash/SPARK-5914 and squashes the following commits:
e288e56 [Judy Nash] Fix spacing and refactor code
1de3c0e [Judy Nash] [SPARK-5914] Enable spark-submit to run requiring only user permission on windows
This metric is incomplete, because the files are memory mapped, so much of the read from disk occurs later as tasks actually read the file's data.
This should be merged into 1.3, so that we never expose this incorrect metric to users.
CC pwendell ksakellis sryza
Author: Kay Ousterhout <kayousterhout@gmail.com>
Closes#4749 from kayousterhout/SPARK-5982 and squashes the following commits:
9737b5e [Kay Ousterhout] More fixes
a1eb300 [Kay Ousterhout] Removed one more use of local read time
cf13497 [Kay Ousterhout] [SPARK-5982] Remove incorrectwq Local Read Time Metric
As documented in createDirectory, the result of createDirectory is not registered for automatic removal. Currently there are 4 directories left in `/tmp` after just running `pyspark`.
Author: Milan Straka <fox@ucw.cz>
Closes#4759 from foxik/remove-tmp-dirs and squashes the following commits:
280450d [Milan Straka] Use createTempDir in getOrCreateLocalRootDirs...
In Standalone mode, the number of cores in Completed Applications of the Master Web Page will always be zero, if sc.stop() is called.
But the number will always be right, if sc.stop() is not called.
The reason maybe:
after sc.stop() is called, the function removeExecutor of class ApplicationInfo will be called, thus reduce the variable coresGranted to zero. The variable coresGranted is used to display the number of Cores on the Web Page.
Author: guliangliang <guliangliang@qiyi.com>
Closes#4567 from marsishandsome/Spark5771 and squashes the following commits:
694796e [guliangliang] remove duplicate code
a20e390 [guliangliang] change to Cores Using & Requested
0c19c95 [guliangliang] change Cores to Cores (max)
cfbd97d [guliangliang] [SPARK-5771] Number of Cores in Completed Applications of Standalone Master Web Page always be 0 if sc.stop() is called
Add Slf4jSink to Spark Metrics using Coda Hale's SlfjReporter.
This sends metrics to log4j, allowing spark users to reuse log4j pipeline for metrics collection.
Reviewed existing unit tests and didn't see any sink-related tests. Please advise on if tests should be added.
Author: Judy <judynash@microsoft.com>
Author: judynash <judynash@microsoft.com>
Closes#4644 from judynash/master and squashes the following commits:
57ef214 [judynash] doc clarification and indent fixes
a751a66 [Judy] Spark-5708: Add Slf4jSink to Spark Metrics
For screenshot see: https://issues.apache.org/jira/browse/SPARK-5965
This was caused by 20a6013106.
Author: Andrew Or <andrew@databricks.com>
Closes#4739 from andrewor14/user-jar-blocker and squashes the following commits:
23c4a9e [Andrew Or] Use right argument
Patch should be self-explanatory
pwendell JoshRosen
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#4741 from tdas/SPARK-5967 and squashes the following commits:
653b5bb [Tathagata Das] Fixed the fix and added test
e2de972 [Tathagata Das] Clear stages which have no corresponding active jobs.
https://issues.apache.org/jira/browse/SPARK-5724
In AkkaUtil, we set several failure detector related the parameters as following
```
al akkaConf = ConfigFactory.parseMap(conf.getAkkaConf.toMap[String, String])
.withFallback(akkaSslConfig).withFallback(ConfigFactory.parseString(
s"""
|akka.daemonic = on
|akka.loggers = [""akka.event.slf4j.Slf4jLogger""]
|akka.stdout-loglevel = "ERROR"
|akka.jvm-exit-on-fatal-error = off
|akka.remote.require-cookie = "$requireCookie"
|akka.remote.secure-cookie = "$secureCookie"
|akka.remote.transport-failure-detector.heartbeat-interval = $akkaHeartBeatInterval s
|akka.remote.transport-failure-detector.acceptable-heartbeat-pause = $akkaHeartBeatPauses s
|akka.remote.transport-failure-detector.threshold = $akkaFailureDetector
|akka.actor.provider = "akka.remote.RemoteActorRefProvider"
|akka.remote.netty.tcp.transport-class = "akka.remote.transport.netty.NettyTransport"
|akka.remote.netty.tcp.hostname = "$host"
|akka.remote.netty.tcp.port = $port
|akka.remote.netty.tcp.tcp-nodelay = on
|akka.remote.netty.tcp.connection-timeout = $akkaTimeout s
|akka.remote.netty.tcp.maximum-frame-size = ${akkaFrameSize}B
|akka.remote.netty.tcp.execution-pool-size = $akkaThreads
|akka.actor.default-dispatcher.throughput = $akkaBatchSize
|akka.log-config-on-start = $logAkkaConfig
|akka.remote.log-remote-lifecycle-events = $lifecycleEvents
|akka.log-dead-letters = $lifecycleEvents
|akka.log-dead-letters-during-shutdown = $lifecycleEvents
""".stripMargin))
```
Actually, we do not have any parameter naming "akka.remote.transport-failure-detector.threshold"
see: http://doc.akka.io/docs/akka/2.3.4/general/configuration.html
what we have is "akka.remote.watch-failure-detector.threshold"
Author: CodingCat <zhunansjtu@gmail.com>
Closes#4512 from CodingCat/SPARK-5724 and squashes the following commits:
bafe56e [CodingCat] fix the grammar in configuration doc
338296e [CodingCat] remove failure-detector related info
8bfcfd4 [CodingCat] fix the misconfiguration in AkkaUtils
Instead of storing a strong reference to accumulators, I've replaced this with a weak reference and updated any code that uses these accumulators to check whether the reference resolves before using the accumulator. A weak reference will be cleared when there is no longer an existing copy of the variable versus using a soft reference in which case accumulators would only be cleared when the GC explicitly ran out of memory.
Author: Ilya Ganelin <ilya.ganelin@capitalone.com>
Closes#4021 from ilganeli/SPARK-3885 and squashes the following commits:
4ba9575 [Ilya Ganelin] Fixed error in test suite
8510943 [Ilya Ganelin] Extra code
bb76ef0 [Ilya Ganelin] File deleted somehow
283a333 [Ilya Ganelin] Added cleanup method for accumulators to remove stale references within Accumulators.original to accumulators that are now out of scope
345fd4f [Ilya Ganelin] Merge remote-tracking branch 'upstream/master' into SPARK-3885
7485a82 [Ilya Ganelin] Fixed build error
c8e0f2b [Ilya Ganelin] Added working test for accumulator garbage collection
94ce754 [Ilya Ganelin] Still not being properly garbage collected
8722b63 [Ilya Ganelin] Fixing gc test
7414a9c [Ilya Ganelin] Added test for accumulator garbage collection
18d62ec [Ilya Ganelin] Updated to throw Exception when accessing a GCd accumulator
9a81928 [Ilya Ganelin] Reverting permissions changes
28f705c [Ilya Ganelin] Merge remote-tracking branch 'upstream/master' into SPARK-3885
b820ab4b [Ilya Ganelin] reset
d78f4bf [Ilya Ganelin] Removed obsolete comment
0746e61 [Ilya Ganelin] Updated DAGSchedulerSUite to fix bug
3350852 [Ilya Ganelin] Updated DAGScheduler and Suite to correctly use new implementation of WeakRef Accumulator storage
c49066a [Ilya Ganelin] Merge remote-tracking branch 'upstream/master' into SPARK-3885
cbb9023 [Ilya Ganelin] Merge remote-tracking branch 'upstream/master' into SPARK-3885
a77d11b [Ilya Ganelin] Updated Accumulators class to store weak references instead of strong references to allow garbage collection of old accumulators
...th RangePartitioner
Author: Aaron Josephs <ajoseph4@binghamton.edu>
Closes#1381 from aaronjosephs/PLAT-911 and squashes the following commits:
e30ade5 [Aaron Josephs] [SPARK-911] allow efficient queries for a range if RDD is partitioned with RangePartitioner
Fix a overflow bug in JdbcRDD when calculating partitions for large BIGINT ids
Author: Evan Yu <ehotou@gmail.com>
Closes#4701 from hotou/SPARK-5860 and squashes the following commits:
9e038d1 [Evan Yu] [SPARK-5860][CORE] Prevent overflowing at the length level
7883ad9 [Evan Yu] [SPARK-5860][CORE] Prevent overflowing at the length level
c88755a [Evan Yu] [SPARK-5860][CORE] switch to BigInt instead of BigDecimal
4e9ff4f [Evan Yu] [SPARK-5860][CORE] JdbcRDD overflow on large range with high number of partitions
Continue to see IllegalStateException in YARN cluster mode. Adding a simple workaround for now.
Author: Nishkam Ravi <nravi@cloudera.com>
Author: nishkamravi2 <nishkamravi@gmail.com>
Author: nravi <nravi@c1704.halxg.cloudera.com>
Closes#4690 from nishkamravi2/master_nravi and squashes the following commits:
d453197 [nishkamravi2] Update NewHadoopRDD.scala
6f41a1d [nishkamravi2] Update NewHadoopRDD.scala
0ce2c32 [nishkamravi2] Update HadoopRDD.scala
f7e33c2 [Nishkam Ravi] Merge branch 'master_nravi' of https://github.com/nishkamravi2/spark into master_nravi
ba1eb8b [Nishkam Ravi] Try-catch block around the two occurrences of removeShutDownHook. Deletion of semi-redundant occurrences of expensive operation inShutDown.
71d0e17 [Nishkam Ravi] Merge branch 'master' of https://github.com/apache/spark into master_nravi
494d8c0 [nishkamravi2] Update DiskBlockManager.scala
3c5ddba [nishkamravi2] Update DiskBlockManager.scala
f0d12de [Nishkam Ravi] Workaround for IllegalStateException caused by recent changes to BlockManager.stop
79ea8b4 [Nishkam Ravi] Merge branch 'master' of https://github.com/apache/spark into master_nravi
b446edc [Nishkam Ravi] Merge branch 'master' of https://github.com/apache/spark into master_nravi
5c9a4cb [nishkamravi2] Update TaskSetManagerSuite.scala
535295a [nishkamravi2] Update TaskSetManager.scala
3e1b616 [Nishkam Ravi] Modify test for maxResultSize
9f6583e [Nishkam Ravi] Changes to maxResultSize code (improve error message and add condition to check if maxResultSize > 0)
5f8f9ed [Nishkam Ravi] Merge branch 'master' of https://github.com/apache/spark into master_nravi
636a9ff [nishkamravi2] Update YarnAllocator.scala
8f76c8b [Nishkam Ravi] Doc change for yarn memory overhead
35daa64 [Nishkam Ravi] Slight change in the doc for yarn memory overhead
5ac2ec1 [Nishkam Ravi] Remove out
dac1047 [Nishkam Ravi] Additional documentation for yarn memory overhead issue
42c2c3d [Nishkam Ravi] Additional changes for yarn memory overhead issue
362da5e [Nishkam Ravi] Additional changes for yarn memory overhead
c726bd9 [Nishkam Ravi] Merge branch 'master' of https://github.com/apache/spark into master_nravi
f00fa31 [Nishkam Ravi] Improving logging for AM memoryOverhead
1cf2d1e [nishkamravi2] Update YarnAllocator.scala
ebcde10 [Nishkam Ravi] Modify default YARN memory_overhead-- from an additive constant to a multiplier (redone to resolve merge conflicts)
2e69f11 [Nishkam Ravi] Merge branch 'master' of https://github.com/apache/spark into master_nravi
efd688a [Nishkam Ravi] Merge branch 'master' of https://github.com/apache/spark
2b630f9 [nravi] Accept memory input as "30g", "512M" instead of an int value, to be consistent with rest of Spark
3bf8fad [nravi] Merge branch 'master' of https://github.com/apache/spark
5423a03 [nravi] Merge branch 'master' of https://github.com/apache/spark
eb663ca [nravi] Merge branch 'master' of https://github.com/apache/spark
df2aeb1 [nravi] Improved fix for ConcurrentModificationIssue (Spark-1097, Hadoop-10456)
6b840f0 [nravi] Undo the fix for SPARK-1758 (the problem is fixed)
5108700 [nravi] Fix in Spark for the Concurrent thread modification issue (SPARK-1097, HADOOP-10456)
681b36f [nravi] Fix for SPARK-1758: failing test org.apache.spark.JavaAPISuite.wholeTextFiles
Follow-on to https://github.com/apache/spark/pull/4591
Document isEmpty / take / parallelize and their interaction with (an empty) RDD[Nothing] and RDD[Null]. Also, fix a marginally related minor issue with histogram() and EmptyRDD.
CC rxin since you reviewed the last one although I imagine this is an uncontroversial resolution.
Author: Sean Owen <sowen@cloudera.com>
Closes#4698 from srowen/SPARK-5744.2 and squashes the following commits:
9b2a811 [Sean Owen] 2 extra javadoc fixes
d1b9fba [Sean Owen] Document isEmpty / take / parallelize and their interaction with (an empty) RDD[Nothing] and RDD[Null]. Also, fix a marginally related minor issue with histogram() and EmptyRDD.
In the general case, Spillable's heuristic of checking for memory stress
on every 32nd item after 1000 items are read is good enough. In general,
we do not want to be enacting the spilling checks until later on in the
job; checking for disk-spilling too early can produce unacceptable
performance impact in trivial cases.
However, there are non-trivial cases, particularly if each serialized
object is large, where checking for the necessity to spill too late
would allow the memory to overflow. Consider if every item is 1.5 MB in
size, and the heap size is 1000 MB. Then clearly if we only try to spill
the in-memory contents to disk after 1000 items are read, we would have
already accumulated 1500 MB of RAM and overflowed the heap.
Patch #3656 attempted to circumvent this by checking the need to spill
on every single item read, but that would cause unacceptable performance
in the general case. However, the convoluted cases above should not be
forced to be refactored to shrink the data items. Therefore it makes
sense that the memory spilling thresholds be configurable.
Author: mcheah <mcheah@palantir.com>
Closes#4420 from mingyukim/memory-spill-configurable and squashes the following commits:
6e2509f [mcheah] [SPARK-4808] Removing minimum number of elements read before spill check
Another one from JoshRosen 's wish list. The first commit is much smaller and removes 2 of the 4 Clock classes. The second is much larger, necessary for consolidating the streaming one. I put together implementations in the way that seemed simplest. Almost all the change is standardizing class and method names.
Author: Sean Owen <sowen@cloudera.com>
Closes#4514 from srowen/SPARK-4682 and squashes the following commits:
5ed3a03 [Sean Owen] Javadoc Clock classes; make ManualClock private[spark]
169dd13 [Sean Owen] Add support for legacy org.apache.spark.streaming clock class names
277785a [Sean Owen] Reduce the net change in this patch by reversing some unnecessary syntax changes along the way
b5e53df [Sean Owen] FakeClock -> ManualClock; getTime() -> getTimeMillis()
160863a [Sean Owen] Consolidate Streaming Clock class into common util Clock
7c956b2 [Sean Owen] Consolidate Clocks except for Streaming Clock
This PR adds a `finalize` method in DiskMapIterator to clean up the resources even if some exception happens during processing data.
Author: zsxwing <zsxwing@gmail.com>
Closes#4219 from zsxwing/SPARK-5423 and squashes the following commits:
d4b2ca6 [zsxwing] Cleanup resources in DiskMapIterator.finalize to ensure deleting the temp file
The stability of the new submission gateway assumes that the arguments in `DriverWrapper` are consistent across multiple Spark versions. However, this is not at all clear from the code itself. In fact, this was broken in 20a6013106, which is fortunately OK because both that commit and the original commit that added this gateway are part of the same release.
To prevent this from happening again we should at the very least add a huge warning where appropriate.
Author: Andrew Or <andrew@databricks.com>
Closes#4687 from andrewor14/driver-wrapper-warning and squashes the following commits:
7989b56 [Andrew Or] Add huge compatibility warning
A variable `shutdownCallback` in SparkDeploySchedulerBackend can be accessed from multiple threads so it should be enclosed by synchronized block.
Author: Kousuke Saruta <sarutak@oss.nttdata.co.jp>
Closes#3781 from sarutak/SPARK-4949 and squashes the following commits:
c146c93 [Kousuke Saruta] Removed "setShutdownCallback" method
c7265dc [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-4949
42ca528 [Kousuke Saruta] Changed the declaration of the variable "shutdownCallback" as a volatile reference instead of AtomicReference
552df7c [Kousuke Saruta] Changed the declaration of the variable "shutdownCallback" as a volatile reference instead of AtomicReference
f556819 [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-4949
1b60fd1 [Kousuke Saruta] Improved the locking logics
5942765 [Kousuke Saruta] Enclosed shutdownCallback in SparkDeploySchedulerBackend by synchronized block
This patch addresses a race condition in DAGScheduler by properly synchronizing accesses to its `cacheLocs` map.
This map is accessed by the `getCacheLocs` and `clearCacheLocs()` methods, which can be called by separate threads, since DAGScheduler's `getPreferredLocs()` method is called by SparkContext and indirectly calls `getCacheLocs()`. If this map is cleared by the DAGScheduler event processing thread while a user thread is submitting a job and computing preferred locations, then this can cause the user thread to throw "NoSuchElementException: key not found" errors.
Most accesses to DAGScheduler's internal state do not need synchronization because that state is only accessed from the event processing loop's thread. An alternative approach to fixing this bug would be to refactor this code so that SparkContext sends the DAGScheduler a message in order to get the list of preferred locations. However, this would involve more extensive changes to this code and would be significantly harder to backport to maintenance branches since some of the related code has undergone significant refactoring (e.g. the introduction of EventLoop). Since `cacheLocs` is the only state that's accessed in this way, adding simple synchronization seems like a better short-term fix.
See #3345 for additional context.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#4660 from JoshRosen/SPARK-4454 and squashes the following commits:
12d64ba [Josh Rosen] Properly synchronize accesses to DAGScheduler cacheLocs map.
Currently, PySpark does not support narrow dependency during cogroup/join when the two RDDs have the partitioner, another unnecessary shuffle stage will come in.
The Python implementation of cogroup/join is different than Scala one, it depends on union() and partitionBy(). This patch will try to use PartitionerAwareUnionRDD() in union(), when all the RDDs have the same partitioner. It also fix `reservePartitioner` in all the map() or mapPartitions(), then partitionBy() can skip the unnecessary shuffle stage.
Author: Davies Liu <davies@databricks.com>
Closes#4629 from davies/narrow and squashes the following commits:
dffe34e [Davies Liu] improve test, check number of stages for join/cogroup
1ed3ba2 [Davies Liu] Merge branch 'master' of github.com:apache/spark into narrow
4d29932 [Davies Liu] address comment
cc28d97 [Davies Liu] add unit tests
940245e [Davies Liu] address comments
ff5a0a6 [Davies Liu] skip the partitionBy() on Python side
eb26c62 [Davies Liu] narrow dependency in PySpark
This patch bring the pull based progress API into Python, also a example in Python.
Author: Davies Liu <davies@databricks.com>
Closes#3027 from davies/progress_api and squashes the following commits:
b1ba984 [Davies Liu] fix style
d3b9253 [Davies Liu] add tests, mute the exception after stop
4297327 [Davies Liu] Merge branch 'master' of github.com:apache/spark into progress_api
969fa9d [Davies Liu] Merge branch 'master' of github.com:apache/spark into progress_api
25590c9 [Davies Liu] update with Java API
360de2d [Davies Liu] Merge branch 'master' of github.com:apache/spark into progress_api
c0f1021 [Davies Liu] Merge branch 'master' of github.com:apache/spark into progress_api
023afb3 [Davies Liu] add Python API and example for progress API
Avoid call to remove shutdown hook being called from shutdown hook
CC pwendell JoshRosen MattWhelan
Author: Sean Owen <sowen@cloudera.com>
Closes#4648 from srowen/SPARK-5841.2 and squashes the following commits:
51548db [Sean Owen] Avoid call to remove shutdown hook being called from shutdown hook
hasShutdownDeleteTachyonDir(file: TachyonFile) should use shutdownDeleteTachyonPaths(not shutdownDeletePaths) to determine Whether contain file. To solve it ,delete two unused function.
Author: xukun 00228947 <xukun.xu@huawei.com>
Author: viper-kun <xukun.xu@huawei.com>
Closes#4418 from viper-kun/deleteunusedfun and squashes the following commits:
87340eb [viper-kun] fix style
3d6c69e [xukun 00228947] fix bug
2bc397e [xukun 00228947] deleteunusedfun
previous behavior was to log an error; this is fine in the general
case where no `spark.metrics.conf` parameter was specified, in which
case a default `metrics.properties` is looked for, and the execption
logged and suppressed if it doesn't exist.
if the user has purposefully specified a metrics.conf file, however,
it makes more sense to show them an error when said file doesn't
exist.
Author: Ryan Williams <ryan.blake.williams@gmail.com>
Closes#4571 from ryan-williams/metrics and squashes the following commits:
5bccb14 [Ryan Williams] private-ize some MetricsConfig members
08ff998 [Ryan Williams] rename METRICS_CONF: DEFAULT_METRICS_CONF_FILENAME
f4d7fab [Ryan Williams] fix tests
ad24b0e [Ryan Williams] add "metrics.properties" to .rat-excludes
94e810b [Ryan Williams] throw if nonexistent Sink class is specified
31d2c30 [Ryan Williams] metrics code review feedback
56287db [Ryan Williams] throw if nonexistent metrics config file provided
There is chance of dead lock that the Python process is waiting for ending mark from JVM, but which is eaten by corrupted stream.
This PR checks the ending mark from Python in non-block way, so it will not blocked by Python process.
There is a small chance that the ending mark is sent by Python process but not available right now, then Python process will not be used.
cc JoshRosen pwendell
Author: Davies Liu <davies@databricks.com>
Closes#4601 from davies/freeze and squashes the following commits:
e15a8c3 [Davies Liu] update logging
890329c [Davies Liu] Merge branch 'freeze' of github.com:davies/spark into freeze
2bd2228 [Davies Liu] add more logging
656d544 [Davies Liu] Update PythonRDD.scala
05e1085 [Davies Liu] check ending mark in non-block way
This patch improves SubmitRestProtocol's handling of invalid JSON requests in cases where those requests were parsable as JSON but not as JSON objects (e.g. they could be parsed as ararys or strings). I replaced an unchecked cast with pattern-matching and added a new test case.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#4637 from JoshRosen/rest-protocol-cast and squashes the following commits:
b3f282b [Josh Rosen] [SPARK-5849] Handle more types of invalid JSON in SubmitRestProtocolMessage.parseAction
The exception in Python writer thread will shutdown executor.
Author: Davies Liu <davies@databricks.com>
Closes#4577 from davies/exception and squashes the following commits:
eb0ceff [Davies Liu] Update PythonRDD.scala
139b0db [Davies Liu] capture the exception in python write thread
The timer is a GC root, and failing to terminate it leaks SparkContext
instances.
Author: Matt Whelan <mwhelan@perka.com>
Closes#4635 from MattWhelan/SPARK-5848 and squashes the following commits:
2a1e8a5 [Matt Whelan] SPARK-5848: teardown the ConsoleProgressBar timer
This patch changes PySpark so that the GatewayServer's port is communicated back to the Python process that launches it over a local socket instead of a pipe. The old pipe-based approach was brittle and could fail if `spark-submit` printed unexpected to stdout.
To accomplish this, I wrote a custom `PythonGatewayServer.main()` function to use in place of Py4J's `GatewayServer.main()`.
Closes#3424.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#4603 from JoshRosen/SPARK-2313 and squashes the following commits:
6a7740b [Josh Rosen] Remove EchoOutputThread since it's no longer needed
0db501f [Josh Rosen] Use select() so that we don't block if GatewayServer dies.
9bdb4b6 [Josh Rosen] Handle case where getListeningPort returns -1
3fb7ed1 [Josh Rosen] Remove stdout=PIPE
2458934 [Josh Rosen] Use underscore to mark env var. as private
d12c95d [Josh Rosen] Use Logging and Utils.tryOrExit()
e5f9730 [Josh Rosen] Wrap everything in a giant try-block
2f70689 [Josh Rosen] Use stdin PIPE to share fate with driver
8bf956e [Josh Rosen] Initial cut at passing Py4J gateway port back to driver via socket
After a call to stop, the shutdown hook is redundant, and causes a
memory leak.
Author: Matt Whelan <mwhelan@perka.com>
Closes#4627 from MattWhelan/SPARK-5841 and squashes the following commits:
d5f5c7f [Matt Whelan] SPARK-5841: remove DiskBlockManager shutdown hook on stop
Author: Ryan Williams <ryan.blake.williams@gmail.com>
Closes#4573 from ryan-williams/history and squashes the following commits:
a8647ec [Ryan Williams] fix test calls to .replay()
98aa3fe [Ryan Williams] include filename in history-parsing error message
8deecf0 [Ryan Williams] add line number to history-parsing error message
b668b52 [Ryan Williams] add log info line to history-eventlog parsing
Naturally, we may need to add an option to print the spark version in spark script. It is pretty common in script tool.
Author: uncleGen <hustyugm@gmail.com>
Author: genmao.ygm <genmao.ygm@alibaba-inc.com>
Closes#4522 from uncleGen/master-clean-150211 and squashes the following commits:
9f2127c [genmao.ygm] revert the behavior of "-v"
015ddee [uncleGen] minor changes
463f02c [uncleGen] minor changes
The changes are summarized in the commit message. Test or test-related code accounts for 90% of the lines changed.
Author: Andrew Or <andrew@databricks.com>
Closes#4557 from andrewor14/rest-tests and squashes the following commits:
b4dc980 [Andrew Or] Merge branch 'master' of github.com:apache/spark into rest-tests
b55e40f [Andrew Or] Add test for unknown fields
cc96993 [Andrew Or] private[spark] -> private[rest]
578cf45 [Andrew Or] Clean up test code a little
d82d971 [Andrew Or] v1 -> serverVersion
ea48f65 [Andrew Or] Merge branch 'master' of github.com:apache/spark into rest-tests
00999a8 [Andrew Or] Revamp tests + fix a few corner cases
mateiz was excluding the time to write this final file from the shuffle write time intentional?
Author: Kay Ousterhout <kayousterhout@gmail.com>
Closes#4559 from kayousterhout/SPARK-5762 and squashes the following commits:
5c6f3d9 [Kay Ousterhout] Use foreach
94e4237 [Kay Ousterhout] Removed open time metrics added inadvertently
ace156c [Kay Ousterhout] Moved metrics to finally block
d773276 [Kay Ousterhout] Use nano time
5a59906 [Kay Ousterhout] [SPARK-5762] Fix shuffle write time for sort-based shuffle
ksakellis I stumbled on your JIRA for this yesterday; I know it's assigned to you but I'd already done this for my own uses a while ago so thought I could help save you the work of doing it! Hopefully this doesn't duplicate any work you've already done.
Here's a screenshot of what the UI looks like:
Based on a discussion with pwendell, I put the data read remotely in as an additional metric rather than showing it in brackets as you'd suggested, Kostas. The assumption here is that the average user doesn't care about the differentiation between local / remote data, so it's better not to pollute the UI.
I also added data about the local read time, which I've found very helpful for debugging, but I didn't put it in the UI because I think it's probably something not a ton of people will need to use.
With this change, the total read time and total write time shown in the UI will be equal, fixing a long-term source of user confusion:
Author: Kay Ousterhout <kayousterhout@gmail.com>
Closes#4510 from kayousterhout/SPARK-5645 and squashes the following commits:
4a0182c [Kay Ousterhout] oops
5f5da1b [Kay Ousterhout] Small style fix
5da04cf [Kay Ousterhout] Addressed more comments from Kostas
ba05149 [Kay Ousterhout] Remove parens
a9dc685 [Kay Ousterhout] Kostas comment, test fix
33d2e2d [Kay Ousterhout] Merge remote-tracking branch 'upstream/master' into SPARK-5645
347e2cd [Kay Ousterhout] [SPARK-5645] Added local read bytes/time to task metrics
[Was previously PR4507]
As per SPARK-5655, recently committed code chmod 700s all application files created on the local fs by a spark executor. This is both unnecessary and broken on YARN, where files created in the nodemanager's working directory are already owned by the user running the job and the 'yarn' group. Group read permission is also needed for the auxiliary shuffle service to be able to read the files, as this is running as the 'yarn' user.
Author: Andrew Rowson <github@growse.com>
Closes#4509 from growse/master and squashes the following commits:
7ca993c [Andrew Rowson] Moved chmod700 functionality into Utils.getOrCreateLocalRootDirs
f57ce6b [Andrew Rowson] [SPARK-5655] Don't chmod700 application files if running in a YARN container
If the user specifies a bad REST URL, the server will throw an NPE instead of propagating the error back. This is because the default `ErrorServlet` has the wrong prefix. This is a one line fix. I am will add more comprehensive tests in a separate patch.
Author: Andrew Or <andrew@databricks.com>
Closes#4518 from andrewor14/rest-npe and squashes the following commits:
16b15bc [Andrew Or] Correct ErrorServlet context prefix
Previously, SparkHadoopWriter always committed its tasks without question. The problem is that when speculation is enabled sometimes this can result in multiple tasks committing their output to the same file. Even though an HDFS-writing task may be re-launched due to speculation, the original task is not killed and may eventually commit as well.
This can cause strange race conditions where multiple tasks that commit interfere with each other, with the result being that some partition files are actually lost entirely. For more context on these kinds of scenarios, see SPARK-4879.
In Hadoop MapReduce jobs, the application master is a central coordinator that authorizes whether or not any given task can commit. Before a task commits its output, it queries the application master as to whether or not such a commit is safe, and the application master does bookkeeping as tasks are requesting commits. Duplicate tasks that would write to files that were already written to from other tasks are prohibited from committing.
This patch emulates that functionality - the crucial missing component was a central arbitrator, which is now a module called the OutputCommitCoordinator. The coordinator lives on the driver and the executors can obtain a reference to this actor and request its permission to commit. As tasks commit and are reported as completed successfully or unsuccessfully by the DAGScheduler, the commit coordinator is informed of the task completion events as well to update its internal state.
Future work includes more rigorous unit testing and extra optimizations should this patch cause a performance regression. It is unclear what the overall cost of communicating back to the driver on every hadoop-committing task will be. It's also important for those hitting this issue to backport this onto previous version of Spark because the bug has serious consequences, that is, data is lost.
Currently, the OutputCommitCoordinator is only used when `spark.speculation` is true. It can be disabled by setting `spark.hadoop.outputCommitCoordination.enabled=false` in SparkConf.
This patch is an updated version of #4155 (by mccheah), which in turn was an updated version of this PR.
Closes#4155.
Author: mcheah <mcheah@palantir.com>
Author: Josh Rosen <joshrosen@databricks.com>
Closes#4066 from JoshRosen/SPARK-4879-sparkhadoopwriter-fix and squashes the following commits:
658116b [Josh Rosen] Merge remote-tracking branch 'origin/master' into SPARK-4879-sparkhadoopwriter-fix
ed783b2 [Josh Rosen] Address Andrew’s feedback.
e7be65a [Josh Rosen] Merge remote-tracking branch 'origin/master' into SPARK-4879-sparkhadoopwriter-fix
14861ea [Josh Rosen] splitID -> partitionID in a few places
ed8b554 [Josh Rosen] Merge remote-tracking branch 'origin/master' into SPARK-4879-sparkhadoopwriter-fix
48d5c1c [Josh Rosen] Roll back copiesRunning change in TaskSetManager
3969f5f [Josh Rosen] Re-enable guarding of commit coordination with spark.speculation setting.
ede7590 [Josh Rosen] Add test to ensure that a job that denies all commits cannot complete successfully.
97da5fe [Josh Rosen] Use actor only for RPC; call methods directly in DAGScheduler.
f582574 [Josh Rosen] Some cleanup in OutputCommitCoordinatorSuite
a7c0e29 [Josh Rosen] Create fake TaskInfo using dummy fields instead of Mockito.
997b41b [Josh Rosen] Roll back unnecessary DAGSchedulerSingleThreadedProcessLoop refactoring:
459310a [Josh Rosen] Roll back TaskSetManager changes that broke other tests.
dd00b7c [Josh Rosen] Move CommitDeniedException to executors package; remove `@DeveloperAPI` annotation.
c79df98 [Josh Rosen] Some misc. code style + doc changes:
f7d69c5 [Josh Rosen] Merge remote-tracking branch 'origin/master' into SPARK-4879-sparkhadoopwriter-fix
92e6dc9 [Josh Rosen] Bug fix: use task ID instead of StageID to index into authorizedCommitters.
b344bad [Josh Rosen] (Temporarily) re-enable “always coordinate” for testing purposes.
0aec91e [Josh Rosen] Only coordinate when speculation is enabled; add configuration option to bypass new coordination.
594e41a [mcheah] Fixing a scalastyle error
60a47f4 [mcheah] Writing proper unit test for OutputCommitCoordinator and fixing bugs.
d63f63f [mcheah] Fixing compiler error
9fe6495 [mcheah] Fixing scalastyle
1df2a91 [mcheah] Throwing exception if SparkHadoopWriter commit denied
d431144 [mcheah] Using more concurrency to process OutputCommitCoordinator requests.
c334255 [mcheah] Properly handling messages that could be sent after actor shutdown.
8d5a091 [mcheah] Was mistakenly serializing the accumulator in test suite.
9c6a4fa [mcheah] More OutputCommitCoordinator cleanup on stop()
78eb1b5 [mcheah] Better OutputCommitCoordinatorActor stopping; simpler canCommit
83de900 [mcheah] Making the OutputCommitCoordinatorMessage serializable
abc7db4 [mcheah] TaskInfo can't be null in DAGSchedulerSuite
f135a8e [mcheah] Moving the output commit coordinator from class into method.
1c2b219 [mcheah] Renaming oudated names for test function classes
66a71cd [mcheah] Removing whitespace modifications
6b543ba [mcheah] Removing redundant accumulator in unit test
c9decc6 [mcheah] Scalastyle fixes
bc80770 [mcheah] Unit tests for OutputCommitCoordinator
6e6f748 [mcheah] [SPARK-4879] Use the Spark driver to authorize Hadoop commits.
Hadoop has a feature that allows users to impersonate other users
when submitting applications or talking to HDFS, for example. These
impersonated users are referred generally as "proxy users".
Services such as Oozie or Hive use this feature to run applications
as the requesting user.
This change makes SparkSubmit accept a new command line option to
run the application as a proxy user. It also fixes the plumbing
of the user name through the UI (and a couple of other places) to
refer to the correct user running the application, which can be
different than `sys.props("user.name")` even without proxies (e.g.
when using kerberos).
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#4405 from vanzin/SPARK-5493 and squashes the following commits:
df82427 [Marcelo Vanzin] Clarify the reason for the special exception handling.
05bfc08 [Marcelo Vanzin] Remove unneeded annotation.
4840de9 [Marcelo Vanzin] Review feedback.
8af06ff [Marcelo Vanzin] Fix usage string.
2e4fa8f [Marcelo Vanzin] Merge branch 'master' into SPARK-5493
b6c947d [Marcelo Vanzin] Merge branch 'master' into SPARK-5493
0540d38 [Marcelo Vanzin] [SPARK-5493] [core] Add option to impersonate user.
